Privacy-Preserving Machine Learning with Homomorphic Encryption among Multi-Parties

Sammanfattning

In the Internet age, with the rapid expansion of data and the widespread integration of artificial intelligence, privacy concerns have intensified and robust privacy-preserving methods are in demand. While non- cryptographic approaches provide a systematic solution to prevent privacy leakage during data mining, others continue to explore the potential of Secure Multi-Party Computation (SMPC) techniques. This thesis addresses a number of challenges in machine learning by developing a framework for privacy-preserving neural networks, using Lattice-Based Cryptography (LBC) and an established Homomorphic Encryption (HE) scheme, called Cheon-Kim-Kim-Song (CKKS). SCHEME, an innovative framework designed for real-world multi-party scenarios with resilience against future quantum computer attacks, is optimized for neural network compatibility and implements a modularized Convolutional Neural Network (CNN) for seamless integration and minimal configuration. The framework demonstrates competitive performance against state-of- the-art HE-based CNN models, and excels in runtime and memory efficiency with multi-threading support. Tested on a subset of the CIFAR-10 dataset, SCHEME achieves up to 85% end-to-end classification accuracy and high precision approximation in individual layers, validating its effectiveness for privacy-preserving image classification tasks. In additional, this work also contributes a comprehensive analysis of privacy-preserving methods, explores mechanisms for group key exchange, discusses security guarantees and evaluates framework robustness. The proposed framework shows promise for applications requiring strong data confidentiality, such as healthcare and finance. By leveraging CKKS-based LBC, this work advances scalable, secure solutions for multi-party machine learning on encrypted data, bringing privacy-preserving methods closer to practical deployment in data-sensitive applications.